harvis-coding

name	harvis-coding
description	Coding tasks delegated to OpenCode. OpenCode uses Kimi K2.5 internally and returns compact JSON summaries (not raw terminal output) - saving 5-10x tokens per task.
read_when	["User asks for a coding task, bug fix, feature, or code review","Working on multi-file changes or architectural refactors"]

Harvis Coding Skill (OpenCode-Powered)

All coding tasks are delegated to OpenCode which runs Kimi K2.5 and returns compact JSON summaries instead of raw terminal output. This saves 5-10x tokens.

The Tool: opencode

Use the opencode tool for ALL coding tasks:

opencode({
  prompt: "Your coding task here. Be specific: what files, goal, constraints.",
  session_id: "optional - from previous call to continue same session"
})

Workflow

1. Delegate to OpenCode — Call opencode with a specific prompt
2. Read the result — You'll get compact JSON with:
   • files_changed: list of files modified
   • commands_run: list of commands executed
   • tests_passed: boolean
   • summary: one sentence summary
   • status: "ok" | "error" | "needs_review"
3. Report to user — Summarize what was done based on the JSON

Continuing a Session

If working on the same codebase across multiple calls, pass session_id from the previous result to continue the same OpenCode session:

opencode({
  prompt: "Continue working on the same task...",
  session_id: "abc123-from-previous-result"
})

Example

User: "Create a new API endpoint for user authentication"

You:

opencode({
  prompt: "Create a new API endpoint for user authentication in python_back_end/main.py. Add POST /api/auth/login that accepts email/password, validates against the database, and returns a JWT token. Include proper error handling for invalid credentials.",
})

Then read the JSON result and report:

• "Created POST /api/auth/login endpoint in main.py with JWT authentication"
• Files changed: ["python_back_end/main.py"]
• Tests passed: true/false

Why OpenCode?

Before (raw bash)	After (OpenCode tool)
500-2000 tokens/step	50-100 tokens/step
Full terminal output in context	Compact JSON only

OpenCode is already configured with:

• Model: Kimi K2.5
• Agent: harvis-coder (compact output mode)
• Server: localhost:4096 (in same pod)